Human Fall Detection In Indoor Scenes Based On Deep Learning

With an aging population increasingly serious situation,the problem of the elderly fall more and more attention.Falling can easily cause coma,disability,or loss of life in the elderly.Falling has become the leading killer of elderly health problems.The current fall detection method has the following problems.One is that it is difficult to extract features with good expressive ability to distinguish many actions that are similar to falls.Second,the accuracy and real-time nature of fall event detection cannot be well balanced.Based on deep learning,this article explores and analyzes the fall behavior of the elderly in daily life.The main research content includes three aspects: human target detection,feature extraction and fall detection.The main research contents of this article are as follows.In terms of target detection,this paper analyzes and compares the mainstream target detection algorithms,improves and optimizes the Faster-RCNN framework,and increases the resolution of feature maps by fusing feature maps of multiple convolutional layers,improving the accuracy of target detection rate.The test results on the fall data set show that this method improves the detection accuracy and recall rate.In terms of video classification,the mainstream 3D convolution and dual-stream network two different behavior recognition methods are used for research.Based on the 3D convolutional neural network,the spatiotemporal features of motion information representing the fall action are extracted,and robust action features are obtained.This paper proposes a fall detection model based on OVR(one-versus-rest)SVMs.The detection model uses equalinterval image features for fusion to effectively recognize falling actions.The time segmentation network of the Inception structure extracts the spatial features representing the appearance characteristics of the fall action and the time features representing the movement characteristics of the fall action,and uses the Softmax classifier to construct the fall model,which implements a multi-feature fusion mechanism based on the decision-making layer.The classification probability of the branch is estimated to be averaged according to the corresponding category,which further improves the accuracy of fall detection.Recognize falling actions.In addition,with the development of posture estimation algorithms in recent years,different from the perspective of extracting the characteristics of different behavior categories from the network,posture estimation is to directly detect and return to the position of important key points of the human body in the image or video.Based on this,this article first analyzes the current status of pose estimation,and introduces the OpenPose algorithm in detail.After obtaining the key points of the human body,it is sent to the SVM behavior classifier,and the specific positions of the obtained key points of the human body are directly used as the characteristic of judging behavior so as to detect falling behavior.